基于 A-DResUnet 的语音增强方法

为了更精确地从语谱图中提取特征信息,提出了一种基于 A-DResUnet 的语音增强方法。 A-DResUnet 模型在 ResUnet 模型的基础上融合了空洞卷积,提升捕获语音上下文信息的能力;同时在编码器中加入卷积注意力模块(CBAM),提高对噪声 谱图特征的关注。 实验结果表明,与模型输出目标为干净语音语谱图相比,用噪声谱图作为模型输出目标时,该模型对未知噪 声具有更强的分离能力;相较 ResUnet 模型,提出的 A-DResUnet 模型减少了语音细节信息的损失;对比基于 DNN、GAN 的语音 增强方法,PESQ 平均提升了 22. 81%、33. 11%,STOI 平均提升了 9. 62%、15. 33%,为复杂环境下的语音增强提供了一种更有效 的方法。

Speech enhancement method based on A-DResUnet

In order to extract feature information from spectrogram more accurately, this paper proposes a speech enhancement method based on A-DResUnet ( attention-dilated ResUnet). The A-DResUnet model incorporates dilated convolution on the basis of ResUnet model to improve the ability to capture the contextual information of speech; at the same time, the convolution block attention module (CBAM) is added into the ResUnet encoder to improve the attention to the features of the noise spectrogram. The experimental results show that when the noise spectrum is used as the output target of the model, the model has a stronger ability to separate unknown noise than when the output target of the model is clean speech spectrum; compared with the ResUnet model, the proposed A-DResUnet model reduces the loss of speech detail information; compared with the speech enhancement methods based on DNN and GAN, PESQ increased by an average of 22. 81%, 33. 11%, STOI increased by an average of 9. 62%, 15. 33%, which is a more effective method for speech enhancement in complex environments.